Discussion AMD's Future CPU-APU Gone ARM !!!

[NTMBK]

Qualcomm plans to put Oryon cores everywhere. They're just starting with laptops first because they see a clear opportunity where AMD/Intel have failed to compete with Apple Silicon.

Next, they'll put Oryon on phones. Then servers.

Bloomberg - Are you a robot?

Qualcomm's Oryon cores will come to the Snapdragon 8 Gen 4

Qualcomm has announced that its computing Oryon cores will be coming to mobile in 2024.

www.xda-developers.com

Oh come on... Qualcomm had an internal server chip, complete with custom CPU core, then shut the whole thing down 5 years ago. They then proceeded to buy a server startup, kill their server chip, and refocus them on consumer chips instead.

I'm not going to take any Qualcomm server rumours seriously until they're on the second generation of shipping server chips. Until then, I'll just assume anything is liable to be cancelled again.

 

[Shivansps]

If Zen5 is so good, why does AMD need an ARM APU?

As i said i dont think Zen5C cores is what AMD should use when trying to make something to compite with ADL-N successors and future WOA socs, its too much. They need to go lower power/heat than that and i just dont think they are going to develop a new small core.
The error here is to think that if they do, they are going to go for a X5 for a premium soc, thats not gona happen they have Zen5C for that.

 

[Ghostsonplanets]

If Zen5 is so good, why does AMD need an ARM APU?

As far we know, it's a custom order from a customer (MS). No more, no less.

when trying to make something to compite with ADL-N successors and future WOA socs

That's Sonoma Valley. 4 Z5C on SF4X for cheap value.
I do think AMD will do a new Cat core in the future.

 

[uzzi38]

As i said i dont think Zen5C cores is what AMD should use when trying to make something to compite with ADL-N successors and future WOA socs, its too much. They need to go lower power/heat than that and i just dont think they are going to develop a new small core.
The error here is to think that if they do, they are going to go for a X5 for a premium soc, thats not gona happen they have Zen5C for that.

I think you might be underestimating how performance Zen 5C will likely end up being. Also how efficient the C cores are - they're good enough that at 15W AMD found 2+4 was faster than 6+0 (see: 7540U vs 7545U).

Even assuming it peaks at similar clocks to Zen 4C (around 3.8GHz) and Zen 5 is only a 20% uplift on Zen 4, that's still a quad-core chip with single thread performance akin to Zen 4 running at ~4.5GHz. That's pretty much a baseline for expected performance, I wouldn't be surprised if in certain benchmarks - like for example Geekbench 6 ST - you'd get scores closing in on PHX/HWK ST performance. For devices costing <$600, that's a genuinely good base point to work from.

At that performance level, given the price target I'd rather see a focus on minimising system level power to maximise on battery life, and IMO the way to do that isn't really to focus on a new core, but rather to focus on uncore power. That's where both AMD and Intel are furthest behind ARM vendors, their expertise in mobile helps them much more considerably there.

 

[Tigerick]

As far we know, it's a custom order from a customer (MS). No more, no less.

That's Sonoma Valley. 4 Z5C on SF4X for cheap value.
I do think AMD will do a new Cat core in the future.

I actually have different opinion about custom chip for Surface. As for the past SoC, MS have been renamed the SoC to their own naming scheme. But we all know the original SoCs are still being sold in the retail market. Just like upcoming Surface X with X-Elite SoC, I am sure MS will custom name it. And we all can buy notebook with X-Eilte in the future.

 

[Tigerick]

Once upon a time, they had XScales, so who knows ?

I also remembered Intel sold XScale and focus on x86. I believe Intel rather helping NV makes ARM SoC than making their own...So Intel is really the last OEM in the world to make their own ARM SoC.

At the same time, Intel has been trying to offload legacy in their x86 in hope to reduce core footprint and better compete with ARM.

 

[FlameTail]

That's Sonoma Valley. 4 Z5C on SF4X for cheap value

S4X sounds like a bad choice of a node for a low power chip. S4X is extremely leaky.

 

[Tigerick]

S4X sounds like a bad choice of a node for a low power chip. S4X is extremely leaky.

And that's why AMD choose Sonoma Valley for SF4X cause Zen5c cores are not designed to be high clock speed as explained by @uzzi38 ...

 

[Ghostsonplanets]

I actually have different opinion about custom chip for Surface. As for the past SoC, MS have been renamed the SoC to their own naming scheme. But we all know the original SoCs are still being sold in the retail market. Just like upcoming Surface X with X-Elite SoC, I am sure MS will custom name it. And we all can buy notebook with X-Eilte in the future.

It's a good point raised. But I'm unsure if AMD would want to commercialize an Arm APU. It certainly would be interesting, though.

S4X sounds like a bad choice of a node for a low power chip. S4X is extremely leaky.

https://twitter.com/x/status/1655655896213598214

Samsung claimed a 23% reduction in power. Though it didn't clarify if it was against SF4P (4LPP+) or something else.

AFAIK S4X should be a node with better v/f and more relaxed pitch to achieve higher clocks. Ideal for HPC applications. But there's too few details to have any certainty.

 

[Shivansps]

I think you might be underestimating how performance Zen 5C will likely end up being. Also how efficient the C cores are - they're good enough that at 15W AMD found 2+4 was faster than 6+0 (see: 7540U vs 7545U).

Even assuming it peaks at similar clocks to Zen 4C (around 3.8GHz) and Zen 5 is only a 20% uplift on Zen 4, that's still a quad-core chip with single thread performance akin to Zen 4 running at ~4.5GHz. That's pretty much a baseline for expected performance, I wouldn't be surprised if in certain benchmarks - like for example Geekbench 6 ST - you'd get scores closing in on PHX/HWK ST performance. For devices costing <$600, that's a genuinely good base point to work from.

At that performance level, given the price target I'd rather see a focus on minimising system level power to maximise on battery life, and IMO the way to do that isn't really to focus on a new core, but rather to focus on uncore power. That's where both AMD and Intel are furthest behind ARM vendors, their expertise in mobile helps them much more considerably there.

Thats exactly what im saying, Zen5C will be too fast for the job, on this sector you dont need that much CPU power. If a quad Zen5C can get anywhere near a "quad" PHX in cpu performance then it is too fast.
Competition around that area will be mainly 4x/x6, A78, x4 A710/A715 or x2/x4 ADL-N(or the successor)... Anything bigger will be just too expensive and well into mainstream ryzen territory.

 

[Ghostsonplanets]

Do we know if AMD and Samsung license agreements allow for Samsung to use their RDNA Exynos on Laptops? I ask because I saw some Samsung related leakers claiming that Samsung will use the Exynos 2500 on a Galaxy Book, too.

If it's true, I assume it will be the successor the Galaxy Book 4 Edge with Snap X Elite.

 

[Tigerick]

Do we know if AMD and Samsung license agreements allow for Samsung to use their RDNA Exynos on Laptops? I ask because I saw some Samsung related leakers claiming that Samsung will use the Exynos 2500 on a Galaxy Book, too.

If it's true, I assume it will be the successor the Galaxy Book 4 Edge with Snap X Elite.

Source about E2500 on GalaxyBook?

 

[NostaSeronx]

They need to go lower power/heat than that and i just dont think they are going to develop a new small core.

Standard core option <-- Been here since Zen1
Dense core option <-- Been here since Zen4
Low Power core option <-- Will be added with Zen5

The reason to avoid ARM is the new product clause:
Post-Mendocino w/ AMD64 => $100+ soc -> No fees for AMD64
Post-Mendocino w/ ARM64 => $100+ soc -> $600+ product times 3% ARM64 royalty tax -> $18+ per product times 10 million = $180+ million in royalties

AMD64 can be used for lower costing solutions, while ARM64 can't be used. However, the low-cost solution brought on by Cat cores require cheap nodes. So, a Zen5E/LP core on 4nm will not be as cheap as a synthesized AMD64 core on 12FDX.

Lowest average cost range:
A9-9400 = $80-$120 average / 8GB+128GB
3015e/Ce = $120-200 average / 4GB+64GB
Silver 7120u = $200-300 average / 4GB+128GB
3 7320U = $300-400 average / 8GB+256/512GB

Stoney/Pollock is still more ideal than Mendocino dual-core and quad-core in the cost department.

Competition around that area will be mainly 4x/x6, A78, x4 A710/A715 or x2/x4 ADL-N(or the successor)... Anything bigger will be just too expensive and well into mainstream ryzen territory.

Should be noted that RISC-V is basically going to nuke this.

RISC-V $150 gives 16-cores of A78 on 12nm :: TDP 30W down to 5W
ARM $120 gives 4-cores of A78 on 6nm and 4xA76 on 16nm :: TDP 8W/12W
Intel $120 gives 4-cores of Zen+ on Intel 7 :: TDP 6W

The closest option from AMD is cheapo 4300U/5300U at $180/$190 respectively. Since, Mendocino products aren't cheap or affordable. Quad-core Mendocino happens to be near double the cost of Quad-core Renoir/Cezanne. I doubt SV will be affordable at all, since AMD's track record at TSMC has been awful. The salvaged big die is cheaper than the base small die.

Expense of supply chain:
South Korea = Highest
Taiwan = 2nd place (AMD has never gotten affordability with TSMC :: Crowned Eagle/Bhavani was purely fabbed at Dresden for example)
America = 3rd place (Pollock/Raven2 never got as cheap as Germany/Singapore sourced products)
Germany = 4th place (Stoney was real cheap)
Singapore = 5th place (Sempron 90nm/65nm was really cheap)

 

[Ghostsonplanets]

Source about E2500 on GalaxyBook?

This person:

https://twitter.com/x/status/1775724470256419218

Exynos 2500 A: 8C in a 1+5+2*
Exynos 2500 B: 10C in 2+6+2*

*If I'm not mistaken

They do seem to be more or less reputable with Samsung leaks. I think they were the first to claim the Exynos 2500 core configuration.

 

[poke01]

dupe

 

[poke01]

Apple is a walled garden and that tight integretion allows them to extract performance at the cost of freedom of choice.

Can we stop saying this? Apple's M chips are not good only because of the walled garden but because they actually are good.
Linux runs bare metal on Mx Macs and you get same CPU performance on arm64 Linux as you do on macOS.

Honestly this reads like an ARM/Apple advert. Apple increase market share by 100% in three years?

Yeah, does he think Apple gives out free cmputers? lol

 

[SpudLobby]

Do we know if AMD and Samsung license agreements allow for Samsung to use their RDNA Exynos on Laptops? I ask because I saw some Samsung related leakers claiming that Samsung will use the Exynos 2500 on a Galaxy Book, too.

If it's true, I assume it will be the successor the Galaxy Book 4 Edge with Snap X Elite.

No way, IMO that Exynos will use Mali.

 

[SpudLobby]

S4X sounds like a bad choice of a node for a low power chip. S4X is extremely leaky.

Where did you see that?

SF4’s nodes have been improved a good bit and with that comes reduction in leakage, it is a variant process of course for HPC but not clear that’s the case the way N4 -> N4X is. To me it looks more like an N5 - N4P sized jump, there’s more than just insane leakage for higher voltages

Like it’s -20-30% power or +10-15% perf over 4LPP, so.

I am sure leakage is higher than the latest 4LPP+ though

 

[soresu]

Oh come on... Qualcomm had an internal server chip, complete with custom CPU core, then shut the whole thing down 5 years ago. They then proceeded to buy a server startup, kill their server chip, and refocus them on consumer chips instead.

I'm not going to take any Qualcomm server rumours seriously until they're on the second generation of shipping server chips. Until then, I'll just assume anything is liable to be cancelled again.

I'm not sure that they cancelled Nuvia's planned server µArch so much as mandated to the engineering team that it was to target consumer markets first, and therefore not to focus on server specific or friendly features (virtualization etc) to begin with.

Also about their previous server CPU core Falkor....

I'm only guessing here, but pretty sure it was an evolution of the OG v8-A SD820 Kryo core, as it seems like they would have tried to at least reuse that R&D rather than create a whole new µArch from scratch.

 

[Mahboi]

I think you might be underestimating how performance Zen 5C will likely end up being. Also how efficient the C cores are - they're good enough that at 15W AMD found 2+4 was faster than 6+0 (see: 7540U vs 7545U).

Hilly shot. No wonder they're killing off Z4 entirely...

At that performance level, given the price target I'd rather see a focus on minimising system level power to maximise on battery life, and IMO the way to do that isn't really to focus on a new core, but rather to focus on uncore power. That's where both AMD and Intel are furthest behind ARM vendors, their expertise in mobile helps them much more considerably there.

Which is what the grapevine says Z6 will do.
It's funny how it works out...or will work out.

 

[Doug S]

Yeah, does he think Apple gives out free cmputers? lol

That's what some people were saying by extrapolating first year M1 sales growth, and was the source of the claims that Apple is eating into Wintel market share and the following claim that Intel/AMD "needed to respond". It was just an initial surge because once Apple's long rumored ARM plans were confirmed, everyone in the market for a Mac was going to wait rather than buying the last generation x86 products.

I have long maintained that Apple operates in its own market and what Intel and AMD (and Qualcomm, when selling Wintel products) has very little effect on them. People make the "Mac vs PC" or for that matter "iPhone vs Android" decision FIRST, then they determine what product they are buying out of the category they have decided upon. There just aren't that many people who are so operating system agnostic that they will cross shop Macbook with LG Gram or iPhone 15 Pro Max with Galaxy S24 Ultra.

 

[FlameTail]

Can we stop saying this? Apple's M chips are not good only because of the walled garden but because they actually are good.

The most efficient CPU in the world.

 

[Thibsie]

The most efficient CPU in the world.

Can we stop marketing BS, please ?
This is not a sale chanel, huh ?

 

[FlameTail]

Can we stop marketing BS, please ?
This is not a sale chanel, huh ?

Is it not the truth?

 

[Tup3x]

As far we know, it's a custom order from a customer (MS). No more, no less.

That's partly the thing. In the past it was either Intel or AMD and those two could basically do what ever they wanted since they held MS hostage (thanks to their duopoly). Now things are about to change - MS has can pretty much pick anyone they want.

Also if other manufactures suddenly start flooding the market with all kinds of ARM machines, it doesn't really matter if AMD have really great x86 cores. I'm pretty sure that x86 is going to lose the PC market eventually. Intel and AMD aren't going to maintain their vendor lock duopoly forever.

That being said, I think AMD could be flexible and play both fields. If it looks like x86 is going to be obsolete, then they could easily give it the finishing blow. Intel alone is not going to last.